Download - TR-NL-4 May 71 NOTE 59p
DOCUMENT RESUME
ED 052 604 EM 009 055
AUTHOR Melton, T. R.TITLE IT1: An Interpretive Tutor.INSTITUTION Texas Univ., Austin. Computer-Assisted Instruction
Lab.SPONS AGENCY National Science Foundation, Washington, D.C.REPORT NO TR-NL-4PUB DATE May 71NOTE 59p.
EDRS PRICEDESCRIPTORS
IDENTIFIERS
ABSTRACT
EDRS Price MF-$0.65 HC-$3.29*Computational Linguistics, *Computer AssistedInstruction, *Computer Programs, Generative Grammar,Language Instruction, Lexicology, *ProgramDescriptions, *Reading Instruction, Semantics,Structural Grammar, Syntax*Interpretive Tutor 1, IT1
A computer-assisted instruction system, called IT1(Interpretive Tutor), is described which is intended to assist astudent's efforts to learn the content of textual material and toevaluate his efforts towarf. that goal. The text is representedinternally in the form of semantic networks with auxiliary structureswhich relate network nodes to particular sentences of Cie originaltext. The data base is completed with the addition of a lexicon, alsoin network form, which is used to store definitions and work-relatedsyntactic information. A very simple command language is used by thestudent to request explanations, definitions, or quizzes over aspecified set of sentences in the original text. A syntax-directed,natural-language sentence generator derives from the data base a setof simple sentences which preserve the meaning of the text passagespecified by the student. The sentences may also be transformed intoquestions which help the student to evaluate his understanding of thetext. A generative grammar and auxiliary functions were developedwhich produce sentences conforming to standard English practice withlimited use of third-person pronouns. A conversational system wasimplemented in the LISP programing language. Potential applicationsof the existing system to language instruction are described. (Author)
U.S. DEPARTMENT OfHEALTH, EDUCATION r, WELFARE
OFFICE OF EDUCATION
THIS DOCOMENT HASBEEN REPRODUCED EXACTLY
AS PICEIVED FROM THEPERSON DR ORGANIZATIONORIGINATING II. POINTS OF YIN OR OPINIONSSTATED DO NOT NECESSARILY
REPRESENT OFFICIAL OFFICE OF EDUCATIONPDSIIIONOR POLICY.
Ill: AN INTERPRETIVE TUTORC:,
CN4
TECHNICAL REPORT NO. NL-4LC\C:C=3
LAIT. R. Melton
May 1971
NATURAL LANGUAGE RESEARCH FOR COMPUTER-ASSISTED INSTRUCTION
Supported By:
THE NATIONAL SCIENCE FOUNDATIONGrant al 509 X
Department of Computer Sciences
and
Computer-Assisted Instruction Laboratory
The University of TexasAustin, Texas
AbztAdet
A CA/ zyztem duckibed which intended to be cased to ZuppoAtthe 4tudent'4 e65okt's to Zeann the content as textuae matekiat and toevataate hi4 own ptogne64 towatd that goal. The text i6 teptezentedinteknatty in the Sown o s zemantic netwothis with auxitiaAy 4tkactukezwhich ketate netwokh nodu to pakticutak 4entence6 as the okiginat text.The data base ins compeeted with the addition of a Zexicon, at/so in netwokh.40AM, which i6 toed to .atone deliinitionh and woad-ketated zyntactic in0Ama-tion. A very 4simpZe command tanguage cased by the 4tudent to tequeztexptanation4, wand-deliinitiono, on quizza (melt a oeciiiied zet a6 zentence6£n the oltiginat text. A zyntax-dikected natukae-Zanguage zentence gene/utakde4ivez Otom the data baze a zet o6 zimpZe zentence6 which pkezekve themeaning o6 the text pa44age 4peci6ied by the 4tudent. Thme 4entence6may be titanzlioamed into quation6 by random zetection as wand's to beomitted on rep aced to produce, keoectively, "Iiitt.-in-the-btanie at tkue-6aae quatiou.
A generative grammar and auxitiaky liunction4 were developedwhich produce 4entence6 which conlimm to 4tandakd Engtizh practice withUnited Lae os thikd-pennon pkonoun6. A convekzationme 4y4tem webs impee-mented in U.T. LISP 1.5.9. Patentiat appticationz as tetativety zimpZeextenaion4 o6 the extating 4y4stem to tanguage inztAuction axe descItibed.
CONTENTS
INTRODUCTION 1
PURPOSE, GOALS, AND LIMITATIONS OF INTERPRETIVE TUTOR 1 . 5
OVERVIEW OF THE SYSTEM 11
THE TUTORIAL SYSTEM: SUPERVISOR AND CONTROL FUNCTIONS . 16
THE DATA BASE: DISCOURSE STRUCTURE AND LEXICON 21
THE GENERATIVE SYSTEM 33
The Syntax Di/acted GenehatoxPxonominatizationDete&mimaion ContentIntetaction with Cont4oZ Function4Noduction o QyzAtion4
SUMMARY AND CONCLUSIONS 49
ImptementationReaatt4Computationat ReciounceoAppticationA and Future Development
APPENDIX A 54
REFERENCES 55
LIST OF TABLES
Table Page
1 Discourse Attributes 26
2 Lexical Attributes Used in Generation 31
3 Lexical Attributes Used in Tutorial System Functions. 31
Figure
1
2
3
LIST OF FIGURES
Interpretive Tutor
Sample Protocol
Generalized Block Diagram of IT1
Page
2
12
18
4A A Simple Discourse Network: Semantic Network . . 22
4B A Simple Discourse Network: Equivalent List Representation 22
4C A Simple Discourse Network: Text 22
5 Conjoined and Prepositional Discourse Nodes . . 27
6 Lexical Entries 32
7 Inflection of Pronouns 43
8 Nominative Noun Phrase Derivation 43
iv
INTRODUCTION
Applying the term Intetpketive Tatox, Simmons (1970) described
a type of computer- assisted instruction (CAI) system which serves as a
student aid in reviewing and promoting understanding of textual material.
In order to accomplish this, the system includes the following components:
(1) Di4coutse StAuctute: a formalized representation of thecontent of the text.
(2) Sentence Genets ton: produces English sentences 'from a speci-fied segment of the text by using the discourse structureas a data base.
(3) Lexicon: provides linguistic information to support thegeneration process and, also, contains definitions of thewords in the text which may be displayed to the studenton demand.
(4) Tatohia Executive: interprets student commands and controlscommunications between student and system. It also super-vises execution of the tutorial functions and carries outthe usual housekeeping activities required in CAI systems.
The interactions among student, instructor, course designer,
and the Interpretive Tutor are illustrated in Figure 1. The student
would be assigned a te':t to study which he would first read and then
he would use the Interpretive Tutor to evaluate his retention of its
content. If any of the text was difficult to read, the generative capacity
of the system could be used to produce "explanations" consisting of
simpler sentences paraphrasing the original text. The lexicon would
provide definitions of words as used in the text. Quizzes might be
generated by appropriate transformations of generated sentences. With
these tools at his disposal the student could conduct a self-directed,
self-motivated review of the text, repeating it until he had satisfied
his own or instructor-imposed criteria of mastery. The successful
1
5
AssignmentsEvaluationReviewInstruction
S
T
U
DE
NT
INSTRUCT&R -USER
Command
41111MM=MMIlliPKolanatipms
Definitions
quizzes
Feedback
2
Performance Summary
ALIME1=7INTERPRETIVE
TUTOR
E SX UE BC SU YT SI T
EE M
GENERATOR
DATA BASE
Response
SUBJECT TextCONSULTANT
LINGUISTICS
CONSULTANT
NATURALLANGUAGECOMPILER
D Ni es tc wo ou r
k
e
41111,11111n. EEEEE
0n
DESIGNER -0, USER
FIGURE 1, -- Interpretive Tutor
3
implementation of such a system and its useful applications to CAI problems
depend upon two important contributions from computational linguistics.
The first is a mechanical aid to the preparation of the discourse structure.
Manual preparation of these formal representations of text is a laborious
and tedious clerical task. Simmons (1971) has developed a natural-language
compiler which is capable of relieving human analysts of much of this
labor. ret, the analysis of natural language is not a task which can
currently be entirely delegated to the computer, nor will it be for
some time to come. However, an interactive version of this compiler
is presently being developed. This will be used by the human analyst
to perform the bulk of the analytical task, requiring only that he render
aid in the translation of passages which are beyond the limitations
of the automated system. It seems that the task of course preparation
could be reduced to reasonable proportions by the use of such a system.
The second contribution from computational linguistics is a
sentence generator of sufficient power to produce an acceptable variety
of paraphrases. While existing systems are by no means capable of
the expressive range of standard English, no such generative virtuosity
is required for the task at hand. The purpose of the generative system
is to simplify and clarify the text, rather than to produce impressive
prose. Therefore, relatively simply syntactic constructions should
serve the purpose of the Interpretive Tutor. Simmons and Slocum (1970)
have developed one generator with sufficient power to produce acceptable
sentences; another is described in this paper.
Complete fulfillment of the interpretive function requires a
third contribution from computational linguistics and from artificial
7
4
intelligence. This would be a semantic component to permit interpretation
of text or student responses. It would also generate questions and
exercises in terms of the implications as well as the surface content
of English sentences. Simmons (1968) experimented with a CAI system
based on semantic analysis of student responses but-concluded that a
semantic component of sufficient generality could riot be implemented
practically today.
Full implementation of the Interpretive Tutor is a goal which,
like all ideals, may never be reached; however, it may be approached
step-by-step from any starting point. The purpose of this paper is
to describe one step: an experimental CAI system, Interpretive Tutor 1,
based on the technology of computational linguistics which might be
practically applied today. The limitations of this initial system are
subsequently discussed, after which some applications and simple extensions
of the basic system will be proposed.
PURPOSE, GOALS, AND LIMITATIONSOF INTERPRETIVE TUTOR 1
Interpretive Tutor 1 (IT1) is a CAI system, but it is functionally
based in computational linguistics. The experimental focus emphasizes
the application of linguistic technology to instructional problems rather
than the refinement of technology in either parent discipline. The
purposes of IT1 are:
(1) To develop basic interactions between tutorial and linguisticcomponents.
(2) To provide a foundation for estimating costs of implementationand operation of applied systems and for evaluating potentialcontributions of existing linguistic technology to educational
problems.
To be useful as a basic instructional tool for a variety of topics,
the structure of IT1 must admit considerable individual variety in
discourse representation and syntactic expression in generated explanations
and questions. This requirement was met by selecting a syntax-directed
generator as well as a flexible structure for the discourse and lexical
data base. These components are discussed later in detail.
IT1 is regarded as the experimental precursor of a family of
similar but more relevant systems. A modular structure of program and
data base is necessary to facilitate experimental modifications and
extensions of the basic system.
The basic function of such Interpretive Tutor systems is to
serve as a remedial and evaluatory tool in the general tutorial process.
Thus they are not "tutorial CAI systems" in the sense of the term used
by Zinn (1968). He referred to a tutorial system that controlled the
9
6
sequence of presentation of instructional material with a greater or
lesser degree of individualized t "eatment according to the performance
of a particular student. In applications of Interpretive Tutor, the
human instructor and the student are expected to carry out the individualizing
and frame-sequencing functions which consequently are omitted from the
basic IT1 system.
Because of the experimental nature of IT1, several typical CAI
functions have not yet been implemented. These include recording latency,
responses, and scores for individual students and response-editing procedures
for detecting misspelled words and similar types of trivial or mechanical
student errors. Such features are essential when students are involved
in either course development, evaluation, or educational operations.
However, the examples of existing CAI systems provide an adequate basis
for cost and design data and the present experimental character of
IT1 does not warrant their implementation.
The linguistic limitations of IT1 result more from the technological
state of computational linguistics and economics than from restriction
of purpose. A practical goal is sought. The means to that goal must
not only be technically demonstrated, but the community of proposed
users must also be able to afford those means. Simmons (1968) has
concluded that, for economic if not technical reasons, the semantic
analysis component must be set aside for the present. Without a semantic
component, syntactic recognition of constructed responses is irrelevant.
Hence, in the present system students are restricted to one-word responses.
The absence of the semantic component also imposes some restrictions
on the performance of the sentence generator. It is clear that the
10
7
"explanations" generated by ITl must be meaning-preserving paraphrases
of the original text. It is neither necessary nor possible to reflect
all human aspects of meaning in machine-generated sentences. Without
attempting to define "meaning," the aspects of meaning which are of
interest are illustrated by example and the meaning that IT1 can preserve
in paraphrases is shown.
Consider the following family of paraphrases:
(1) The old man ate a fish.
(2) The old man did eat a fish.
(3) A fish was eaten by the old man.
(4) . . . the eating of a fish by the old man . .
(5) The old man consumed a fish.
(6) (a) The old man took up pieces of a fish, chewed, andswallowed them.
(b) He continued in this way until the fish was gone.(c) His belly then contained the masticated fish.
Sentence 1 is a straightforward statement describing an event
in the universe external to the speaker and listener. Sentence 2 describes
the same event, but the introduction of the emphatic modal "did" reveals
some attitude of the speaker toward the event he describes or, perhaps,
toward a skeptical listener. At any rate, the speaker's emphasis adds
no information concerning the state of the universe external to the
speaker-listener system.
Sentence 3 also describes the same event, but the speaker has
chosen to emphasize the fish by choosing to mention it as the surface
subject of the sentence. In 4, the entire event has been subsumed as
a noun phrase which is embedded in a larger sentence. These choices
may reveal an attitude of the speaker toward the event, or they may
11
8
have been influenced by previous subject choices in the larger context
in which the sentence appears. In either case no information about
the external universe is added to the content of the first sentence.
There is no equivalent in IT1 to human attitude nor can IT1
make a spontaneous choice among 1, 2, 3, or 4. However, it can produce
any of these paraphrases if appropriate cues are stored in the discourse
structure and if appropriate grammar rules are provided. The original
s.aoject choices are not identified in the discourse structure, and the
author's attitude may be lost in generated sentences.
Sentences 1 through 4 might be regarded as purely "syntactic"
paraphrases in that no content words were added to the standard description
of the event which they all describe. Similarly, sentence 5 is almost
a syntactic paraphrase since there is a sense of "consume" which is
synonymous with the sense of "eat" used in the example. Given appropriate
lexical data, IT1 might produce sentence 5, but the risk of introducing
spurious connotations seems to outweigh any possible benefit.
In contrast to sentences 1 through 5, sentences 6a, b, and c
definitely comprise a "semantic" paraphrase. Whereas 1 through 5 present
a face-value description of the event, 6 mentions several implications
of it. Instead of simply declaring that the old man ate, 6a relates
in some detail what it means to eat, and 6c mentions one of the usual
consequences of eating. A fully interpretive system would permit the
generation of 6a, b, and c under appropriate circumstances, but this
is precisely the kind of operation which requires the semantic component
which is not now available.
12
9
The generative power of IT1 is limited to syntactic paraphrases
such as sentences 1 through 4. English permits many syntactic patterns
within these limitations. If the language is defined to include sentences
which are intelligible, without regard for conformation to any standard
grammatical practice, the set of meaning-preserving paraphrases of an
English sentence may be quite large. However, the basic purpose of Ill
requires that some minimal level of grammatical quality be observed.
The informative function of Ill would be compromised if the student
were obliged to labor over "explanations" which did not conform to common
usage and if the explanations did, by their very clumsiness, attract
his attention to their syntax rather than to their content. Moreover,
the systematic exposure of presumably impressionable students to substandard
English usage is to be avoided.
The IT1 generator was consequently designed to conform to standard
English usage at the level of simple sentences. Subject-verb number-
agreement is enforced, nouns are inflected to indicate number and verbs
for tense. Limited pronominalization is permitted. These features
may be compromised in applications if it is necessary to reduce computational
and storage requirements, but the basic capacity exists.
No extensive evaluation has been made of the ability of the
present generator to produce compound or complex sentences. An experimental
precursor of the present IT1 generator was reasonably successful in
embedding subordinate clauses in complex sentences; however, that system
was applied only to small discourse structures equivalent to two or three
simple sentences. Generation of complex sentences from a large data
base, such as that used in IT1, appears to be significantly governed
13'
10
by a sort of super-grammar which is more closely related to rhetorical
practice than to the syntactic principles that apply to single sentences.
This rhetorical grammar is not adequately understood for general computational
applications. As a result, the generative grammars used in IT1 have
been limited to production of simple sentences. This restriction does
not seem to compromise the basic explanatory and question-generating
purpose of the system.
14
OVERVIEW OF THE SYSTEM
Before describing IT1, it is appropriate to introduce the system
by an example. The general process of application was given in broad
outline in Figure 1. A subject expert prepares or selects a text to
provide the data base. Each sentence in the text is given an identifying
sentence-tag, such as Sl, S2, etc. The text is then translated into
the formal representation of the discourse structure with (or without)
the assistance of a natural language compiler and a linguistics consultant.
The relationship between each sentence of the original text and the
corresponding elements of the discourse structure is preserved by means
of the sentence-tags. As the translation of the text continues, each
word is entered in the lexicon. This lexicon is used to store dictionary
definitions for the words used in the text and to store word-associated
data required by the generator. Together, the lexicon and discourse
structure form the IT1 data base. The addition of a generative grammar
to be used in sentence generation completes the preparation of an IT1
data structure.
The tutorial operation of IT1 is demonstrated in the protocol
of Figure 2. The student has read the assigned text but has had some
difficulty with it. Perhaps the syntax was unclear or there was doubt
concerning the exact meaning of some word used in one of the sentences.
If the student wants to review these troublesome points before undertaking
a quiz, he consults the text and then types
(EXPLAIN S4)
11
15
STOP)))4-<-
(ENTER COMMAND)
STUDENT*STUDENT*MODE****
(EXPLAIN S1)S1(THE PLANTS ARE THE STATELY GRACEFUL COCONUT-PALM)(THEY FIRST APPEAR ON A NEWLY FORMED TROPICAL ISLAND)*END*EXPLAIN***
(EXPLAIN S4)S4
(THE SEED IS PECULIARLY ADAPTED FOR DISTRIBUTION)(OCEAN WAVES AND OCEAN CURRENTS CARRY IT OVER A THOUSAND MILESFROM THE PARENT PLANT)(THEY DISTRIBUTE IT)(IT GROWS ON SOME DISTANT SHORE)*END*EXPLAIN***
(DEFINE COMMERCE S3)(IT1 5)
(WORD IS NOT USED IN SENTENCE)***
(DEFINE COMMERCE S2)COMMERCE(N - THE SALE OR EXCHANGE OF GOODS - TRADE - BUSINESS)***
(EXPLAIN S3)S3(THE SEED *S THICK FIBROUS HUSK AND ITS HARD SHELL SECURELYPROTECT IT FROM THE ACTION OF SEAWATER)*END*EXPLAIN***
(EXPLAIN S2)S2(THE SEED IS THE COCONUT OF COMMERCE)*END*EXPLAIN* * *
FIGURE 2. - -Sample Protocol
1G
12
13
(QUIZ S1 S3)(ANSWER TRUE OR FALSE)(THE PLANTS ARE THE STATELY GRACEFUL COCONUT-PALM)TRUE*RIGHT*(THEY FIRST APPEAR ON A NEWLY FORMED TROPICAL ISLAND)
TRUE*RIGHT*(THE SEED IS THE COCONUT OF COMMERCE)TRUE*RIGHT*(ITS THICK FIBROUS HUSK AND ITS HARD CORE SECURELY PROTECT IT
FROM THE ACTION OF SEAWATER)FALSE*RIGHT*(YOUR SCORE WAS 4)(TOTAL POSSIBLE 4)***
***
(QUIZ S5 S8)(TYPE THE MISSING WORD)(ADAPTATION ACCOUNTS FOR THE SEED *S WIDEDISTRIBUTION*RIGHT*(THE STATELY GRACEFUL COCONUT-PALM ORIGINALLY WAS NATIVE TO
CERTAIN ISLANDS OF THE OR OF TROPICAL AMERICA)
PACIFICWRONG(ANSWER IS TNDIAN*OCEAN)(IT IS NOW FOUND ON MOST TROPICAL SHORES IN THE AND THE
WEST*INDIES)**
(YOUR SCORE WAS 1)(TOTAL POSSIBLE 2)***
STOP
Note.--The notation ** terminates the quiz.
FIGURE 2 (continued).--Sample Protocol
17
14
to command the explanation of the sentence. The al responds by generating
a set of simple sentences which preserve the meaning of the original
sentence, S4.
After requesting other "explanations," the student finds that
the meaning of "commerce" is unclear in the context of a sentence, so
the student enters
(DEFINE COMMERCE S3)
to obtain a definition of the word as used. The system finua that the
word is not used in S3, and displays an appropriate diagnostic response.
The student then enters
(DEFINE COMMERCE S2)
which results in the printing of a dictionary definition for the sense
of the word used in sentence 2.
The student browses through the text briefly and is then ready
for a quiz. He enters
(QUIZ S1 S4)
to limit the scope of the quiz to the first four sentences in the text.
On the basis of a random selection, IT1 decides to use true-false questions
in the quiz and so advises the student. The quiz is then printed, one
question at a time, in the form of simple declarative sentences in each
of which one noun may have been replaced by another noun which renders
the sentence false. After each question has been presented to the
student, IT1 accepts a one-word answer; compares it with the correct
response, and prints an appropriate message. The system accumulates
the student's score as the quiz progresses. After the last question
has been answered, IT1 prints the score and then awaits further instructions.
18
15
The student enters
(QUIZ S5 S8)
to invoke a test over the indicated passage. This time the system elects
to produce fill-in-the-blanks questions, in which one noun is replaced
by a blank. As each response is entered, it is compared with the word
which was omitted, and an appropriate response is generated. Tiring
of the activity, the student enters
to end the quiz, then
STOP
to terminate the session.
This exchange demonstrates the three basic student commands
implemented in this first IT1 system. Some extensions are discussed
later.
In addition to the student mode, IT1 may operate in a supetvison
mode for program testing and data base editing, In this mode the LISP
interpreter is essentially placed at the supervisor's disposal, and
he may invoke any LISP function. Several conversational utility routines,
used in editing the data base, may be entered in this way. The supervisor
may also change program variables or may even define new LISP functions
in the supervisory mode.
19
THE TUTORIAL SUBSYSTEM:SUPERVISOR AND CONTROL FUNCTIONS
The modularity of IT1 is preserved in the tutorial executive
itself. The executive consists of a simple supetvizok, a command-accepting
litinction, and a set of command 6unction4. The basic function of the
supervisor is outlined below.
Each student command consists of a command word and an an.gument
tiat (e.g., QUIZ S1 S5). A command function and a dummy argument list are
associated with each command word. The executive calls a command-accepting
function, ACCEPT-CMD, to acquire the student command and execute preliminary
validity tests. If the command is accepted, the associated command function
name is retrieved, and the real arguments are substituted for the dummy
arguments. The command function is then executed. It supervises its own
student communications, tests for valid arguments, and returns control to the
executive when finished. The present command list is as follows:
Command Command Function
(EXPLAIN Si) EXPLAIN*
(QUIZ Si Sj) QUIZ*
(DEFINE word Si) DEFINE*
STOP STOP
The command function STOP is a special case used to terminate a
session. Eventually, a STOP function will be added which will create and
file a summary of the student's performance during the session, but this
feature has not been implemented in IT1.
16
20
17
The interactions among student, instructor, and the various
components of ITl are summarized in Figure 3. The control functions are
discussed below. Certain details of their interaction with the generator
and data base will necessarily remain unclear until those components have
been described in more detail.
In the following discussion, the term u6a refers to a course
designer, programmer, or other persons who may modify the program or its
data structure. The term 4tudent has the usual meaning of one who is being
taught and, in particular, one whose interaction with the system is limited
to the student commands listed above.
The Command EXPLAIN Si. EXPLAIN* takes as an argument the alleged
sentence-tag specified by the student in the original command. If the
argument is not in fact a sentence-tag, EXPLAIN* prints a diagnostic message
and returns. Otherwise, EXPLAIN* calls the generator to produce the appro-
priate set of sentences. At the time that the text was compiled, the sentence-
tag was used to preserve the relationship between the original sentence and
the corresponding objects in the discourse structure. This was done by
binding the attribute HDVB to each sentence-tag. The attribute HDVB
indicates the object in the discourse structure which was created from the
head verb in the original sentence. This binding is retrieved by EXPLAIN*
which then calls the generator to produce a sentence from the indicated
object. The sentence is returned to EXPLAIN* as the value of the generative
function. The example in Figure 2 (see pg. 12) shows that the explanation of
a text sentence may require the generation of several simple sentences. To
avoid confusion, the term generated paa6age, or simply pasaage, will be
applied to the set of sentences generated by a single EXPLAIN* command. The
21
FIGURE 3, -- Generalized BlocA OlaTinm of IT1.
Revisions
18
E XECUTIVES UBSYSTEM
MASTER CONTROL
USER'nstructo
EDITINGSentences & Data
Tutorial Supervision
Cumwands
CONTROL
D ATAB ASE
go ik TUTORIAL:,SUPERVIIS 0 R
EXPLAIN*
SentencesMsentenrpg
Questions
Response
Feedback Score
Definitions
QUIZ*
11 QuestionOS Sentences
I StatusIlTransformationr.
IT1Answer
PP Processing
Ill
:Ji:Pi.p
DISCOURSESTRUCTURE
LEXICON
WCEIZCIIIMMIEN Generated StringsInternal Strings
Other Data Transfers
Control
19
initial sentence of the generated passage is printed by EXPLAIN* which then
determines (by means discussed later) whether the last possible sentence has
been generated. If not, the generator is called to produce the next sentence
of the passage, and EXPLAIN* prints it. This process continues until the
passage is complete.
The Command DEFINE. DEFINE* is given two arguments corresponding
to the word and sentence-tag specified by the student in the original
command. A dictionary article for the word is retrieved from the lexicon
and printed.
The Command QUIZ Si Si. QUIZ* determines whether both sentence-tags
are legitimate. If they are not, a diagnostic is printed. If both are
legitimate, QUIZ* generates and scores a quiz corresponding to that segment
of the text bounded by Si and Sj (Si = Sj is permitted, and Sj may precede
Si in the text). Each question is a transformation of an ordinary generated
sentence. Basically, QUIZ* produces a passage which is the union of all the
passages which EXPLAIN* would produce for each of the text sentences Si,
Si + 1, . . . , Sj. Answers are accepted and processed after each question
is printed.
Questions are produced by simple substitutions in the original
generated sentence. The basic procedure is to generate a sentence in the
usual way and, then, to substitute for some word in the sentence a word
which renders the sentence false, or a "blank" (i.e., ) to produce
true-false or fill-in-the-blank questions, respectively. A set of true-
false or blanked questions is randomly selected; the user may control the
frequency of choice. All questions in each quiz are of the same type.
23
20
The user may specify a phiohi the syntactic class of words which
are to be candidates for substitution. Nouns were selected in the example
of Figure 2. Any noun in the sentence may be replaced. Only one word
is substituted in each sentence, and a particular word for substitution is
selected at random. Thus, if there are four nouns in a sentence, four
different blank questions may be derived from it. Falsification is randomly
controlled, with a probability of 0.5 that falsification will occur in any
given question. Thus, if there are two possible falsification substituents
for each of the four nouns, the sentence may be transformed into one of
eight possible false statements, and nine different true-false questions may
be derived from the base sentence.
As each substitution is made, the correct answer is stored. In
blank questions, the original word is stored; in the true-false questions,
the truth value )uLe or liatse is stored. The transformed sentence is printed,
and student response is accepted. If the response is a double asterisk (**),
the quiz is terminated. Otherwise, the response is compared with the stored
answer, and an appropriate message is displayed to the student. The total
of incorrect answers is accumulated, and after the last question, the score
is printed.
24
THE DATA BASE:DISCOURSE STRUCTURE AND LEXICON
Three kinds of data have been mentioned which must be provided
in each IT1 system: the di4couroe AtAuctuke, the texicon, and a gtammak.
The discourse structure and lexicon form one continuous structure, which
will be referred to as the data base. The grammar is a separate structure
containing rules which govern the derivation of sentences from the data
base. In the following section, the grammar and generative algorithm
are discussed in more detail, and the nature of their correspondence
becomes clear. The general structure of the data base and the content
chosen in this experimental implementation are described presently.
The structure of the data base is based on Quillian's (1966)
notion of a 6emantie netwania, wherein each cognitive entity in a complex
unit of meaning is defined by relations which exist between it and other
cognitive entities. Such structures may be represented by a graph sucn
as that of Figure 4a. Each entity is represented by a labeled node,
and relations by labeled arcs. Simmons (1970a) has shown that the attribute-
value lists of Figure 4b are equivalent to the graphical representation.
They provide a convenient machine-readable representation for computational
applications.
The node G1 represents an event, namely the eating of a fish
by an old man. The fact that the event in question was a specific taken
of the larger class of events identified by the verb eat is specified
by the TOK relation between G1 and the lexical entry eat. The AGT arc
indicates the active participant in the event, G2, the man, and OBJ
21
25
22
mmmmm -rya orom
AGT*
TOK
TOK*
DET
A . -- Semantic Network
(G1
(G2
(G3
(G4
(TOK eat ) (AGT G2) .(OBJ G4) (AUX PAST) )
(TOK man) (MOD G3) (AGT* Gl) (DET the) (NBR SING) )
(TOK Old) (MOD,* G2) (DEG POS) )
(TOK fish) (OBJ* Gl) (DET a) (NBR SING) )
, -- Equivalent List Representation
S The old man ate a fish.
C -- Text
FIGURE 4, -- A Simple Discourse Network
2G
23
indicates the participant receiving the action. The inverse relations
TOK*, AGT*, etc., are represented in Figure 4a by a broken line. Lexical
entries are represented in lower case letters, and node add arc labels
are in upper case letters to distinguish them from lexical entries.
The arcs NBR and AUX are not true relations in the sense of
the arcs mentioned above, since they do not indicate objects in the
discourse structure or lexicon. The NBR arc indicates the syntactic
number of an object, and AUX indicates information required to produce
the appropriate inflection ate of the verb eat. These and similar
relations are termed 6eatate4. Relations and features are collectively
termed attkibute6.
Each lexical entry is the head of a similar structure which
defines that entry. The lexical substructures are omitted for the
present.
The following conventions are used in reference to discourse
networks:
(1) The node G1 is said to dominate G2, G4, and G3, since Gi
is superior to the others in the hierarchy.
(2) G2 is the value of the attribute AGT in Gl; therefore,
G2 is said to be dominated by the relation AGT in G1 or,
simply, AGT dominates G2 if the identity of the dominant
node is of no interest.
(3) A node is said to be headed by the lexical entry indicated
by the TOK of the node. Thus, G1 is headed by eat. Similarly,
G1 is said to be verb - headed since its head is a verb.
27
24
The structure of Fiaures 4a and 4b was derived from the sentence
in 4c, and might have been derived from any of the examples 1-6 of page 7 in
the preceding section. It could have been derived mechanically from
any of sentences 1-5, and this is the basis of the discourse structure
used in IT1. Semantic networks are used as canonical forms from which
a variety of English representations might be generated in a systematic
way, so that a human derives approximately the same meaning from any
of the representations. The operational notion of "meaning" is defined
in terms of lexical content and syntactic form. If a given English
string X is well formed and meaningful to a competent reader, then it
defines a semantic equivalence class of strings. A different string Y
is in the class of X if:
(1) All the content words in both X and Y are morphological
derivatives of common lexical roots.
(2) The syntactic construction of Y preserves the meaning of X.
In the absence of any effective definition of "meaning" it is
necessary to illustrate the intent of the syntactic restriction by
an example. Consider the following strings, all of which might be derived
from any one of the set:
(1) The man ate the fish.
(2) The fish was eaten by the man.
(3) . . . the eating of the fish by the man . .
(4) The fish ate the man.
Strings 1-3 are in the same semantic class. String 4 is not in that
class, even though its lexical content is identical with that of the
others. The syntactic form of 4 leads to quite a different interpretation.
23
25
In 3, the "man-eat-fish" idea is subsumed by a larger sentence,
while it is given independent syntactic status as a sentence in 1 and 2.
Thus semantic class membership is partially independent of surface
syntactic classification.
The generative process takes place in a two - dimensional environment
in which the lexical content of a derived string must be preserved while
the commonly accepted forms of English are imposed upon the string
in such a way that the derived string remains in the semantic class
of its base in the original text. The operational function of the data
base is to provide sufficient information to permit the definition of
effective procedures for generation of sentences which conform to these
restrictions.
The structure presented here is based on that used by Simmons
(1970). Part of its content was suggested by the work of Fillmore (1969)
concerning the content of the lexical component of linguistic systems.
The attributes used in the IT1 discourse structure are defined
in Table 1. The remarks indicate the application of each attribute as
well as the types of syntactic phrases which are derived from them.
The derivations are discussed in a later section. Examples of the data
base used in the first implementation of IT1 are given in Appendix A.
In the present system, conjoined nodes are represented by a
node which is headed by a conjunction which dominates the conjoined
elements by the relations Al and A2. An example (X2) is presented in
Figure 5.
Prepositions may be represented in two ways. In general, locative,
translational, and similar prepositions are represented explicitly in
29
26
TABLE 1
DISCOURSE ATTRIBUTES
Name Value Type Remarks
TOK Lexical Entry See page 21.
AGT, OBJ Discourse NodeDAT, INSTR
Indicate nominal objects which have,respectively, the agentive, objective,or instrumental relations to a verb-headed node.
NOM Discourse Node Indicate nominative, noun-complement,PNOM or adjective-complement of to be.
PADJ
TIME Discourse Node Modifiers of verb-headed nodes, speci-
MANN fying time, manner, and location, re-
LOC spectively. LOC and TIME usuallydominate a preposition-headed node;MANN may dominate adverb- or preposition-headed nodes.
MOD
PMOD
POBJ
Discourse Node
Discourse Node
Discourse Node
Al, A2 Discourse Node
DET Lexical Entry
NBR Number Indicator
AUX Tense Indicator
DEG Degree Indicatorfor Adjective
Adjectual or adverbial modifiers whichare not represented by any explicitrelational attribute.
Indicates a relation normally expressedby the preposition which heads the nodedominated by PMOD.
Dominates the object of a preposition-headed node.
Indicates co-nodes of conjoined nodes.
Indicates a determining article, etc.
S = Singular; PL = Plural.
PR = Present; PAST = past; compoundtenses are constructed syntactically.
POS = Positive; COMP = Comparative;SUPR = Superlative.
30
27
Jack and Jill ran up the hill.
(X1 (TOK run) (AGT X2) (LOC X5) (AUX PAST))
(X2 (TOK and) (NBR PL) (Al X3) (A2 X4) (AGT* X1))
(X3 (TOK Jack) (Al* X2))
(X4 (TOK Jill) (A2* X2))
1X5 (TOK up) (POBJ X6) (LOC* X1))
(X6 (TOK hill) (DET the) (NBR S) (POBJ* X5))
FIGURE 5, -- Conjoined and Prepositional Discourse Nodes.
,31
28
the discourse network by a node such as X5 (see Figure 5), which is
dominated by LOC and headed by a preposition. Other such nodes are
dominated by PMOD. On the other hand, the prepositions by, oi, and others
which are optionally used to represent relations between verbal and
nominal objects, are not represented explicitly. Each such relation
label is given the attribute RELPREP, which indicates the print image
of the appropriate preposition. In generation, a function retrieves
the preposition when it is required in some syntactic environment.
In practice the alternate representations are equivalent in the sense
that a prepositional phrase or a simple noun phrase may be derived from
either.
The Lexicon
The lexicon supports both generative and tutorial operations.
Lexical data are used to supply syntactic information and the actual
print-images for words to be used in generated sentences. The lexicon
also includes definitions for use by the DEFINE* function and falsification
substitutes for use in production of True-False questions.
The lexicon consists of a collection of lexical entries. An
entry is composed of a head and its associated attributes and their
values. Each entry corresponds to one sense of an English word. For
example, the word plant might be used to mean any one of the following:
plant 0: n; a living organism of the vegetable kingdom.
plant 1: n; a factory or mill.
plant 2: vt; to bury a seed or cutting in the ground for growth.
Such a collection of entries, all of which are represented by
the same word, is represented in the lexicon in a manner suggested by the
8°
29
example. One entry is headed by the word itself (the Oth entry). Each
of the remaining entries is headed by a symbol coined by concatenating
a digit with the word, e.g., pleantl. It is this 4ub4cAipted symbol
which is indicated by the TOK relation in the discourse structure and which
is used by the generator as an entry into the lexicon for terminal
operations. Each subscripted word is linked to the parent unsubscripted
word by the attribute SENSES and its inverse, SENSES*, as indicated below:
(plant (SENSES (plant plantl plantl)) (SENSES* plant) . . . )
(plantl (SENSES* plant) . . . )
(plant2 (SENSES* plant) . . . )
The SENSES* attribute is used to fetch the unsubscripted form as a print
image for all entries. Each entry contains the set of attributes and
values required for the computational applications. Since IT1 lacks
semantic operations such as question-answering and analytic capacity,
the linguistic function of the IT1 lexicon is restricted to the storage
of syntactic features and morphological data in the form of pkint-.images
which correspond to the various inflections of the root word which heads
each entry. These inflected forms are used, for example, to indicate
the number of a noun or to obey the rule of number agreement between
a verb and its surface subject.
The usual parts attributed to each verb are: infinitive, present,
past, past participle, and present participle. The third-person singular
is also explicitly specified. Verbs which are irregular in the formation
of the plural third-person past-tense have that form explicitly specified
also and are marked with the flag IREG.. Some verbs, such as appear and
adapt, are closely related to nouns (appearance and adaptation) which
are used in place of the usual gerund in some circumstances. These
33
30
nouns are also associated with the verb. The attribute-names used to
indicate these print images are tabulated in Table 2. The head itself
is the infinitive, and hence, no explicit specification of this form
is required.
Noun entries are headed by the singular form and the plural
form is indicated by the attribute PLF. The genitive is synthesized
by the generator from the singular by adding the ending. Adjective
entries are headed by the positive form, with comparative and superlative
indicated by COMP and SUPR, respectively. Adverb entries are provided
independently of adjectival roots. Prepositions, conjunctions, articles,
and pronwns are either represented directly in the discourse structure
or are synthesized in the generation process and are not included in
the initial lexicon. Several additional attributes are used by the
tutorial subsystem. These attributes and the application of each are
given in Table 3. Except when specified otherwise, these attributes
are used in all entries.
The lexicon is stored in a network structure similar to the
discourse structure. Sample entries are given in Figure 6 in the list
notation.
34
31
TABLE2
LEXICAL ATTRIBUTES USED IN GENERATION
Attribute Name Value (Print-Image for)
PRSF Third person singular, present tense
PAST Regular past tense
PRPART Present participle
Past participle
Derived noun
PPART
NOUN
TABLE 3
LEXICAL ATTRIBUTES USED IN TUTORIAL SYSTEM FUNCTIONS
Attribute Value
LEX
USED-IN
A dictionary article for the given word-senseprinted by the executive in response to the
DEFINE command.
A list of the tags of all sentences in which themorpheme is used. Used by the executive todetermine the appropriate definition in respondingto the DEFINE command.
TOK* A list of discourse nodes in which the head isthe value of TOK.
FALS A list of words which may be substituted for thehead in sentences in order to render the sentencefalse. Used only in noun entries in IT1.
30
32
(COCONUT-PALM ((SENSES COCONUT-PALM) (FALS OAK ELM DATE-PALM FIG)(USED-IN S6) (PLF . COCONUT-PALMS) (LEX N; A TALL SLENDER
TROPICAL PALM COCOS NUCIFERA) (TOK* . G3) ( SENSES* .
COCONUT-PALM)))(APPEAR ((SENSES APPEAR) (LEX VI; TO BECOME VISIBLE) (FALS LEAVE)
(PRSF . APPEARS) (PASTF . APPEARED) (PPAPT . APPEARED) (PRPART.
APPEARING) (PRPART . APPEARING) (TOK* . G10) (SENSES* . APPEAR)))
(BE ((SENSES BE BE1) (USED-IN S1 S2 S3) (IREG . T) (LEX V ; THE
SAME AS ONE OF) (PRSF . IS) (PRPART . BEING) (PRPLF . ARE)
(PSPLF . WERE) (PPART . BEEN) (PASF . WAS) (TOK* G100 G70 G1 G20)
(SENSES* . BE)))
(BE1 ((USED-IN S2 S3 S4) (LEX V; TO HAVE THE PROPERTY OF (HISEYES ARE BLUE) IN A STATE OF (IT IS ADAPTED)) (SENSES* . BE)))
(FIND ((SENSES FIND) (USED-IN S6) (FALS LOSE) (LEX VT ; TO COME
UPON BY CHANCE; TO MEET WITH) (PRSF . FINDS) (PASTF . FOUND)
(PRPART . FINDING) (TOK* . G81) (SENSES* . FIND)))
(STATELY ((COMPR . STATLIER) (SUPR . STATLIEST) (SENSES STATELY)(LEX ADO) (TOK* . G7) (SENSES* . STATELY)))
(ISLAND ((SENSES ISLAND) (FALS MOUNTAIN RIVER) (USED-IN S1 S6 S7
S8) (PLF . ISLANDS) (LEX N - A TRACT OF LAND COMPLETELYSURROUNDED BY WATER BUT TOO SMALL TO BE CALLED A CONTINENT)(TOK* G110 G97 G12 G73) (SENSES* . ISLAND)))
FIGURE 6.--Lexica Entries
36
THE GENERATIVE SYSTEM
The Syntax Ditected Genetaton
The tutorial function of the generative system is the derivation
of sets of sentences corresponding to some text segment specified by an
EXPLAIN* or QUIZ* command. These derived sentences must belong to the
semantic class of the specified text segment. Since the original phrase
structure of the text is lost in the discourse network, there is no simple
correspondence between the discourse nodes and text segments. At the same
time the lexical content of every text segment is reflected in some subnet
of the discourse network, so that the lexical content of the text is
accessible to retrieval by well-defined processes. The generative strategy
employed in IT1 is to define generalized derivational procedures which
produce phrases that preserve the meaning of the original text syntacti-
cally, while guaranteeing the retrieval of at least the full lexical con-
tent of the original text. Other processes are imposed on this basic
generative process to restrict the lexical content of a generated passage
to just that of some specified text segment. The control processes will be
discussed after the basic generative system is described.
The generative system is defined in terms of the derivation of a
pimase from some /Loot node in the discourse network. A phrase is recursively
defined to be either
(1) a word
(2) a sequence of phrases
In the latter case, the defined phrase is said to be the patent phrase and
the individual members of the defining sequence are referred to as
33
.37
34
dependent phrases. Similarly, the root of the parent phrase is the parent
root, and the roots of dependent phrases are dependent roots. A dependent
phrase may be derived from the parent root; i.e., every phrase (root) is
its own parent and its own dependent.
The system is syntax-directed in that the derivational process
is defined by a gene/Wive g4amman supplied by the user, which is inter-
preted by a generative algorithm. Syntactic con;truction is described in
the usual manner of phrase-structure grammars. The lexical content of the
root is preserved by dynamic definition of dependent roots in terms of the
attributes of the parent root. The grammar provides for definition of
conditions upon which the selection of syntactic productions is predicated.
For a simple subset of English it suffices to define these conditions in
terms of the attributes of the parent root. The sentence-rule below illus-
trates these properties of the basic rule format.
S [AGT OBJ] (SUBJ AGT)(VP *)(NP OBJ)/
[AGT] (SUBJ AGT)(VP *)/
[OBJ] (SUBJ OBJ)(PASVP *)
The rule defines three alternate productions for a sentence. Each
production is predicated on a condition which is represented by'the square-
bracketed sequence of attribute-names. A condition is true if it is empty
(i.e., [ ]) or if the root has every attribute named in the condition.
Thus, if the sentence-root has both AGT and OBJ attributes, the first pro-
duction is applied; otherwise, if the root has the AGT attribute, the
second production is applied, and so on. If none of the conditions are true
the "null" production is applied and results in generation of the empty
string containing no words.
38
35
Each pr'4uction is a sequence of paiia (x,y), where x is either
an attribute-name or a nonterminal symbol which defines the syntactic
class of a dependent phrase, and y defines the dependent root. If y is
the symbol "*", then the dependent phrase is derived from the parent root;
otherwise y must be an attribute-name and the dependent phrase is derived
from the discourse node which is dominated by the attribute y in the
parent root.
If x is a nonterminal, then the pair is said to be nontamina.
Otherwise, the pair is said to be -ft/mina, and its generative vaue is the
value of the attribute x in the dependent root. For example, the article
and uninflected noun are produced by terminal pairs in a simple class of
noun phrases defined by the rule
NP [ ] (DET *)(TOK *).
Terminal pairs correspond to the initial step in the recursive
definition of a phrase, and nonterminal pairs correspond to the recursive
step. To complete the analogy, the generative value of a nonterminal
pair is the result of concatenating the generative values of dependent
phrases derived from it. That is, if a pair (NP *) were applied to G2 in
the discourse network of Figure 5, the result would be computed in the
following steps:
(NP *) + ((DET *)(TOK *)) + (the (TOK *)) -4- (the man).
This simple NP-rule does not preserve the lexical content of its
root, since it provides no means for expressing adjectival modifiers or
number, to name just two deficiencies.
39
36
Some power to express transformational operation, synthesis of
function-words, and more complex conditions is necessary for pronominaliza-
tion, subject-verb agreement, and other syntactic properties of the desired
subset of English. This power is provided by adding functions to the
generative system, permitting function-names to appear in grammar rules,
and attaching functions to certain nonterminal symbols. These functions
are classified according to the use which is made of them.
(1) Neudo-iunction4 are attached to nonterminal symbols. They
are used only for their transformational effect on the
discourse network, to set global registers, and the like.
They have no generative value, and their functional value
is ignored by the generative algorithm. A pseudo-function
is attached to the nonterminal SUBJ to retrieve the attribute
NBR and its value from the SUBJ-root, and insert them into the
root node of the parent sentence for later use in enforcing
subject-verb agreement when the verb-phrase is derived from the
nonterminal VP.
(2) ftedicatea are used to augment the basic conditions described
above. An example is PLP which is true if the root has the
attribute-value pair (NBR PL), i.e., if the root is plural,
and otherwise false. Predicates may also have useful side
effects as in the case of MARKF which is described later.
(3) Tamina 6unctionA are used for synthesis of function words,
punctuation, and pronouns, and may also be used for effects
similar to those described in (1) above.
40
37
The application of functions in the grammar and the interpretation
of rules is defined in terms of a generative algorithm. This algorithm
and its auxiliary functions are described in an informal language which
borrows features from both the ALGOL and LISP meta-languages, with some
simple conditions written out in English.
A grammar G is a collection of rules of the form
T C1P1 /C2P2/ /CnPn
where T is a nonterminal symbol. Each Ci is a condition of the form
[cli c2i r ]'m. > 0
where each cji is either an attribute-name or the name of a predicate.
Each production Pi is a sequence of pairs
((xli y2i)(xli y2i) (xki yki)); ki > 0
Wlemeadlx.Ji is either an attribute-name, a nonterminal symbol, or the
name of a terminal function, and each yji is either an attribute-name or
the symbol "*".
The set of rules which compromises a grammar G is defined in the
following manner. Each nonterminal symbol t has the attribute G, and its
value is the sequence of condition-production pairs which defines t in G.
The function gall. is used in the generative process to retrieve the
generative definition of a nonterminal. That is,
get&R;G] = IF t has the attribute G THEN (Cly /CnPn) ELSE nil
The recursive generation process is initially applied to some
specified root node and nonterminal symbol. The appropriate production for
the given root and nonterminal is selected, the successive pairs of the
41
38
production are evaluated, and the results concatenated to produce the
desired string. Production-selection for an arbitrary discourse node
and nonterminal symbol t is carried out by the functions setect, which
is applied to the sequence of condition-production pairs returned by get&
and geond which evaluates conditions.
geond [r; (c1 c2 . . . cm)] =
IF m = 0 THEN true
ELSE IF cl is an attribute name
THEN IF r has the attribute c THEN geond [r; (c2 . . . cm)]
ELSE false
ELSE IF c1[r] THEN geond [r; (c2 . . . cm)]
ELSE false
zetect [r; (C1P1 /C2P2 . . . /CnPn] =
IF n = 0 THEN nil
ELSE IF geond [r; Cl] THEN P1
ELSE zetect [r; (C2P2/ . . . /CnPn)]
The utility function ga4.6oe is used to return the value of an
attribute a in a discourse node A.. For the purpose of this discussion a
discourse node is identified with its defining attribute-value list,
((aivi)(a2v2) (anvn))
The latter, more explicit notation will be used when it can contribute to
clarity.
42
39
gauoc [a; ((aivi)(a2v2) (anvn))] =
IF n = 0 THEN nil
ELSE IF a = al THEN vl
ELSE gazzoc [a; ((a2v2))]
Generative evaluation is carried out by the following set of
recursive functions which evaluate successive sub-elements of productions.
The process is initiated by application of gen1 to an initial root node it
and nonterminal symbol t, where the derivation is governed by grammar G.
genl [r;t] = gen2 [r; Aetect [r; get& [t; G]]]
Next, gen2 transmits successive pairs of the production returned by select
to gen3 for evaluation, and concatenates the resulting dependent phrases.
gen2 Er; ((xlyi)(x2y2) (xnyn))] =
IF n = 0 THEN nil
ELSE cone (gen3 (r; ()WO]; gen2 Ir; ((x2y2) (xnyn))]]
Dependent roots are selected by gen3 which applies .tevaL to
evaluate terminal pairs, or re-initiates the process by applying genl to
nonterminal pairs.
gen3 [r; (x y)] =
IF get& [r; G] THEN
1. x[r;y];
2. gen3:= IF y = "*" THEN genl Er; x]
ELSE genl[gazzoc [y; r]; x]
ELSE IF y = "*" THEN .tevaL [x; y; r]
43 ELSE tevat. [x; y; gazzoc [y; r]]
40
Note that the nonterminal recursive step 2 above is preceded by
application of the pseudo-function attached to x, to the current root r
in step 1, but the value, if any, of the pseudo-function is not appended
to the surface string. If no pseudo-function is attached to x, step 1 has
no effect.
Finally, tevat applies x to the root n and relation y if x is a
terminal function or retrieves its value from the root node if it is an
attribute-name.
tevat [x; y; r] = IF x is a terminal function THEN x[y; r]
ELSE gaiszoc [Y.;
Pumominatization
A brief reference was made earlier to higher-order principles
which govern the derivation of multi-sentence pa'sages from large data bases.
Similar processes govern the natural use of pro ,uns. We are so habituated
to the use of pronouns that passages generated without them seem almost
hopelessly clumsy. Since the minimal goals of the IT1 system might be
compromised by omission of this important detail, a simple mechanism was
implemented which seems to satisfy the minimum requirements.
The present pronominalization mechanism permits the substitut4,on
of pronouns in the third person for modified noun phrases under restricted
conditions based on the content of the current passage. If substitution. is
not permitted, the full noun phrase is produced with all possible modifiers.
44
41
This characteristc leads to tiresome repetition in long passages, but it
is not especially annoying in the IT1 application since the control functions
limit the length of passages to a few sentences. Pronoun substitution makes
three requirements of the generative system:
(1) The eligibility of a discourse node for pronominalizationmust be determined.
(2) The surface syntactic case in which the pronoun appears mustbe specified.
(3) A pronoun must be produced which agrees with its antecedentin number and gender and which is inflected according to case.
These operations are carried out in the following way: Provisional
eligibility is established by a pseudo-function bound to the nonterminal
symbol NOUN, which is invariably applied to the root of a noun-phrase
detivation. This pseudo-function adds the attribute PRON to the root. All
noun-phrase rules test for the presence of this attribute. If it is absent,
the full noun phrase is produced; if it is present, a function is executed
to determine the eligibility of the current root for pronominalization and
to attach the appropriate pronoun to the root if the substitution is per-
mitted. Substitution is permitted only if the selected pronoun has not
already been bound to a different antecedent. If substitution is blocked,
the PRON flag is removed from the current node, and the original antecedent
binding is removed from the pronoun which would otherwise have been used in
substitution. The pronoun is thereafter free for substitution for any
suitable noun phrase.
A minor complexity is introduced by the fact that antecedents are
not bound directly to surface pronouns but to a root-form of the pronoun
instead, which is inflected to indicate syntactic case. It is also necessary
45
42
to introduce surface case syntactically during derivation, since the discourse
relations do not necessarily indicate the surface case of generated noun
phrases. Case is introduced by means of three noun-phrase rules: NNP, ONP,
and GNP, corresponding to nominative, objective, and genitive cases,
respectively. Surface subjects are derived from the nonterminal SUB.) and
NNP's;direct and indirect (e.g., dative) objects are generated as ONP's, as
are prepositional objects. The GNP rule is used to produce possessive
pronouns and adjectives. The interrelations of case, syntactic features,
roots, and surface forms of pronouns are given in Figure 7. This table is
represented internally by a network structure. Antecedents are bound to
the ROOT by the attribute ANT. The pronoun-generating functions operate
on this network to determine the root of the prospective pronoun and to find
the current antecedent bound thereto. Each of the three case-preserving
noun-phrase rules calls a separate pronoun-generating function. Following
the naming convention of the noun-phrase rules, these .Functions are termed
NPRON, OPRON, and GPRON. Each of these functions preserves the case
induced by the grammar rule which invokes it.
Figure 8 is an example of the application of these functions in
production of nominative noun phrases or pronouns and, incidentally, an
example of the method of generating conjoined constituents. The NNP rule
tests for the PRON mark; if the mark is present, NPRON is first applied and
then the NNPI-rule is applied. The NNP1-rule tests for the PRON flag again.
If the pronoun substitution was permitted by NPRON, the appropriate pronoun
is attached to the root of the NNP expansion by the attribute PRON, and
becomes the generative value of NNPI. If NPRON blocked the pronoun substi-
tution, the PRON attribute is removed, the PRON condition on the NNP1-rule
4?
43
SyntacticFeatures
0
0
-,-)
-I-)
d0
Case
0>A4..)0a)
4-)
0 0> >-AM PM 00 0U)
El 400z 0
HE* M S he him his his
SH* F S she her hers her
IT* N S it it --- its
TH* N PI, they them theirs their
FIGURE 7 -- Inflection of Pronouns
NNP * [PRON] (NPRON *) (NNP] *)/
(DET *) (MODP *) (NNP Al) (TOK *) (NNP A2) (PP PMOD)/
[AUX] (NPS *)/(] (NNP *)
NNP], + [PRON] (PRON *)/(1 (NNP *)
FIGURE 8 Nominative Noun Phrase Derivation.
47
44
.,will fail, and the NNP-rule is applied again unconditionally. On this second
application of 11P, the [PRON] condition will fail, and the other alternatives
will be applied. The [Al] condition is used to determine whether the root
is a conjoined constituent. Note that the given rule order permits conjoined
constituents to be pronominalized jointly, or each individual co-constituent
may be pronominalized independently. The [AUX] condition in the third
right-hand side is used to determine whether the root is verb-headed. If
it is, a special noun-phrase rule schema headed by NPS is applied to produce
a nominalized sentence. Finally, if all other conditions have failed, an
uninflected noun phrase is produced by application of the NP rule. The rule
schemata for the other syntactic cases are similar, except that in the
genitive case, the noun is inflected by addition of 's in the full noun
phrase.
Detekmadtion Content
The preceding sections have shown how the generative subsystem
controls content and form of strings of and below the order of simple
sentences. The basic definition of the EXPLAIN command implies that several
simple sentences might be required to contain the content of one complex
sentence of the original text. Thus, the system must control content of
passages, which are strings of higher order than simple sentences. The
performance required for the EXPLAIN command is a restricted special case
of the performance required of the generator as a separate entity. That is,
the generator must produce (in response to one initial request to derive a
sentence from a given node in the discourse network) every possible sentence
that can be derived from the structure. This property is predicated on the
4S)
45
requirement that the discourse network be well-formed in the sense that
every node is related to every other node through some sequence of node-
relation-node paths.
The basic requirement in determining content is that some
syntactic object is to be derived from every discourse node that is related
in any way to the root of the initial sentence. Each time a string is
derived, its root is examined to determine whether it is related to any
unsatisfied node, i.e., if the root is directly related to any other node
which was not the root of a substring of the current string. For every
alternate production for expansion of a nonterminal, the set of attributes
which dominate such nodes is, at worst, the complement of the set of
attributes which dominate the roots of substrings in the expansion in
question, with respect to the universe of all possible discourse relations
in a given data base. A 4upptementarty sentence is derived from every
unsatisfied node. Then, if this procedure is applied at each step in deri-
vation and if the grammar is general in the sense that a simple sentence
can be derived from any node in the network, the generator will have the
desired property. It will exhibit the desired behavior if no sentence is
repeated. The MARKF predicate is used to assure this condition to avoid
infinite repetition of a subset of the possible sentences.
Additional restrictions exist. Clearly, derivation of the supple-
mentary sentences cannot be initiated until the current sentence is completed.
Construction of a grammar which has this property and which exhaustively
derives sentences from all possible supplementary nodes is tedious, at
best, and has been done only for elementary data bases. The task was simpli-
fied for the present application by taking advantage of the character of
49
46
the data base and of the simple-. ,ntence grammar and by introducing two
functions to identify potentic.; supplementary sentence-roots and to
generate sentences from them at the appropriate time. The simple-sentence
grammar was constructed to derive sentences from verb-headed nodes. In
particular, it was constructed so as to derive a substring from every node
dominated by a sentence-root through any chain of forward relations (e.g.,
AGT, as opposed to the inverse AGT*). This task is comparatively simple,
since there is an inherent division of discourse nodes into verbal,
nominal, and other classes, each of which is restricted in the types of
attributes possessed by a node belonging to the class. This grammar was
augmented by two additional functions, PUSH and POP, which are used in the
following way:
PUSH is applied in each noun-phrase derivation. By means
of the inverses of verb-noun relations (AGT*, OBJ*, etc.),PUSH collects a list of the nodes which dominated the rootof the noun-phrase by one of the verb-noun relation. A
list of such nodes is then appended to a global list whichis, in effect, a queue of prospective supplementary sentence-roots.
POP is called when the current sentence is complete. It
retrieves the queue and calls the generator to derive asentence from each node in the queue.
Since PUSH is applied to every noun-phrase root and POP is applied to every
sentence, additional supplementary roots are queued and used throughout the
process. This process does not guarantee exhaustion of every well-formed
data base, since only verb-noun relations are considered in the queueing
process, but it suffices for the purposes of Ill.
59
47
Intetaction with Conttot Functions
The preceding section essentially describes the communications
between the control functions and the generator. In IT1, the POP function
is omitted, and its task is performed by EXPLAIN* and QUIZ*. Both of the
control functions carry out essentially the same process. According to
the procedure described earlier, the student initially specifies by means of
a tag the text sentence from which the passage is to be generated. Two
attributes are associated with each tag:
HDVB identifies the discourse node which is headed bythe principal verb of the original sentence.
SCOPE indicates a list of all other verb-headed nodes whichwere created in the reduction of the sentence.
In the process of deriving the initial sentence, PUSH creates the queue of
supplemental sentence-nodes. After displaying the initial sentence, the
control function carries on in the manner of POP, except that sentences are
derived only from nodes which are defined in the SCOPE list of the specified
sentence tag. This device restricts the content of generated passages to
that of the designated text segment, while the exhaustive character of the
generator assures that every possible sentence root is offered. The user
may, in effect, alter the structure of the text by editing the SCOPE defini-
tions for key sentences.
The system permits the use of different grammars for questions and
explanations to be designated by the user when the system is initialized.
The same grammar is used for both operations in the present system.
Ptoduction Qguaion4
The general procedure used to produce questions was described in
the discussion of the tutorial executive subsystem. A question-producing
51
48
function is given a sentence produced by the generator and modifies it in
the manner described to produce a true-false or fill-the-blank question.
In the case of questions, the generator produces a sentence in a special
form. As the process was previously described, the user selects a ptioti
a Us+ of nonterminal symbols from which a lexical item is derived. This
list is ignored during generation of sentences for EXPLAIN, so that sen-
tences are produced as linear strings of words, for example,
(THE SEED IS THE COCONUT OF COMMERCE)
When sentences are produced for QUIZ, the generative function (at the step
described above as gen3) tests each nonterminal to determine whether it is
selected for question-substitution. If it is, then the phrase (always a
single word) derived from it is returned as a sublist. For example, in
the samples given elsewhere, the nonterminal NOUN was selected for substi-
tution, so that the sentence above would be transmitted to QUIZ* as
(THE (SEED) IS THE (COCONUT) OF (COMMERCE))
To form a question, one of the delimited words is selected at random.
Either " " or a "falsification-word" chosen at random from the list
indicated by the attribute FALS in each lexical entry, is substituted for
the selected word, and the resulting string is printed with the delimiters
removed, as in
(THE SEED IS THE COCONUT OF
(THE SEED IS THE COCONUT OF FINANCE)
Thus questions are produced by simple string manipulations. This procedure
is easily implemented and relatively efficient in operation. Its deficiencies
are discussed in the concluding section.
52
SUMMARY AND CONCLUSIONS
Imptementation
The experimental IT1 system was implemented in LISP for the
CDC 6600 system at The University or Texas Computation Center. The
experimental system is completely interpretive; that is, no functions
were hand-coded in assembly language or compiled.
The discourse structure was prepared by hand from a short passage
of expository prose adapted from Compton'A Encyctopedia. The lexicon,
which is restricted in content to that necessary for the sample text,
was also prepared by hand. A grammar was developed to generate the language
of simple sentences required for the application and refined by trial
and error until satisfactory results were obtained for the sample text.
Results
The samples of Figure 2 were taken from conversational protocols
produced by the author. The reader may compare these sentences with
the original text shown in Appendix A. The language generated by IT1
is limited to simple sentences, including adjectival and adverbial modifiers,
prepositional phrases, and the limited pronominalization described above.
Based on experience with the limited experimental data base, the language
seems adequate for the goals of the system. However, certain infelicitous
results should be mentioned.
One example appears in the first EXPLAIN passage, generated
for Sl, in Figure 2. The original sentence was:
One of the first plants to appear on a newly-formed tropical
island is the stately and graceful coconut palm.
to3
50
The EXPLAIN passage produced by IT1 fails to preserve the meaning of
this sentence, which is adorned with several semantic thorns. It is
difficult to decompose the sentence into several "simpler" sentences
which preserve the original meaning concerning class membership and
temporal order of events without resorting to confusing circumlocutions.
It is probably inappropriate to decompose S1 into two explanatory sentences,
and the discourse structure could be coded to produce essentially the
original sentence. Such problems as this are not uncommon in expository
prose, but hopefully may be dealt with ad hoc without seriously compromising
the aims of the system. Automata cannot "explain" everything, and all
tutorial systems will continue to include a human instructor to render
aid when the automatic components fall short.
The "fill-in-the-blank" questions seem adequate. With the addition
of relatively simple answer-processing functions to detect misspellings
and similar near misses, quizzes of this type would be quite satisfactory.
The true-false questions-are less satisfactory. The simple substitution
technique used to produce false questions may lead to the infliction
of an absurdity such as
The seed is the coconut of music.
or a dilemma such as
The seed is the coconut of industry.
upon the student. To avoid these unfortunate results, the user must
exercise a great deal of care in the selection of falsification substitutions
to be attached to each lexical entry. This care, applied over the collection
of-a large lexicon, amounts to a considerable weight of labor. This
human effort might be better expended on development of automatic techniques
54
51
for selection of falsification substitutions similar to those used by
Carbonell (1970).
Computationat Ruoultzu
IT1, like most natural language processing systems, consumes
processor time and storage space at a fairly conspicuous rate. The
IT1 generator, operating interpretively in LISP, requires an average of
approximately 1.2 seconds of central processor time to produce a simple
sentence. This experimental system is, of course, quite inefficient by
practical standards. The generative grammar is in effect a program
which is interpreted by the generative algorithm, which is in turn
interpreted by the LISP system. Optimization of the graiwor and compila-
tion or hand-coding of the generative algorithm would effect a large
reduction in execution time.
Even thougV. the computational efficiency of the generative system
would be intolerable in practical a'plications, it is not prohibitive in
experimental work at the present level of utilization. However, the
experimental utility of the existing system is seriously compromised by
storage requirements. The existing system requires simultaneous central
memory residency of IT1, its entire data base, and the LISP system.
Experimental progress will be accompanied by increases in the size of the
data base, and the limits of the existing conversational system will soon
be reached. At a relatively early point in development it will be
necessary to convert the system to one in which the data base resides
primarily in mass storage, and is drawn pagewise into central memory as
required for current command operations.
52
Appel cationA and Fu tuhe Development
When considering potential applications for IT1 and potentially
fruitful lines of development of its successors, it is useful to consider
the system in the light of its tutorial functions and data structure.
The existing system has two basic tutorial functions: the retrieval of
information (EXPLAIN and DEFINE), and the generation of tests (QUIZ).
The QUIZ function is, computationally, a special case of EXPLAIN. Questions
differ from explanatory sentences in that part of the retrieved information
is withheld from the student and is subsequently compared with his response
during answer-processing. The scope of the information returned to the
student is limited by the command functions by means of the sentence-tags
and SCOPE and HDVB properties which bind elements of the discourse network
to elements of the original text. These properties impose a secondary
structure un the discourse network. This secondary function is used by the
command functions to initiate and regulate the generation of sentences.
Thus there are two dimensions along which applications might
develop. In one case, the information-retrieval facility might be expanded
by implementation of additional command functions which work within the
existing structure. An example would be a "tell-about" command, which,
given a lexical entry, would return a definition of the entry and a sequence
of sentences derived from every discourse node which was related to the
entry. It would not be difficult to implement such a command, using the
TOK* attribute of lexical entries to initiate: a search for prospective
sentence-roots in the discourse network.
On the other hand, additional secondary structures might be imposed
upon the basic discourse network to serve still other purposes. A
56
53
discourse network might be used, for example, as a base for generation of
sentences to demonstrate variations of standard grammatical practice, or to
produce examples of poor practice to be corrected by the student. For
such a purpose, several generative grammars might be stored simultaneously,
one of them being selected by the command functions according to their
needs. The secondary structure in this case might be one which facilitated
selection of discourse nodes of various types, for example, those headed
by intransitive verbs, or by transaction-verbs, or the like. A single data
base might be used for a variety of instructional purposes. One set of
command functions would simply ignore the secondary structures used by the
other command functions.
Foreign language instruction seems to be a particularly attractive
potential application involving both the basic EXPLAIN function and an
extension of secondary structure. The basic IT1 system might be very help-
ful as a reading aid to a student who is beginning to deal with the more
complex syntactic constructions of a foreign language. With extension in
command and answer-processing functions, the basic generative system might
be used to generate drills in verb inflection and the like.
57
54
APPENDIX A
ORIGINAL TEXT
(S1) One of the first plants to appear on a newly formed tropicalisland is the stately and graceful coconut palm.
(S2) The seed is the coconut of commerce.
(S3) It is securely protected from the action of seawater by itsthick fibrous husk and hard shell.
(S4) It is peculiarly adapted for distribution by ocean waves andcurrents, and may be carried a thousand miles from the parent
plant to grow on some distant shor.l.
(S5) This fact accounts for its wide distribution.
(S5) Originally native to certain islands of the Indian Ocean orof tropical America, it is now found on most tropical islandsin the East Indies and the West Indies.
(S7) These trees lend distinction and attractiveness to the
islands.
(S8) To the cultivation of the coconut is largely due the increasingcommerce and rising civilization of many of these remote islands.
55
REFERENCES
Carbonell, J. R. Mixed instinctive man-computer instruc:lor-L: dialog.Unpublished doctoral dissertation, Massachusetts institute ofTechnology, Cambridge, Massachusetts, June, 1970.
Fillmore, Charles J. Types of lexical information. Ohio State University,
1969 (preprint).
Quillian, M. R. Semantic memory. Unpublished doctoral c 3sertation,Carnegie Institute of Technology, Pittsburgh, Pa., 1966.
Simmons, R. F. Linguistic analysis of constructed student responses in
CAI. TTN-86, Computation Center, The University of Texas atAustin, October, 1968.
Simmons, R. F. Some semantic structures for representing English meanings.Technical Report No. NL-1, National Science Foundation, GrantGJ 509 X, Computer Sciences Department and Computer-AssistedInstruction Laboratory, The University of Texas at Austin,November, 1970a.
Simmons, R. F., & Slocum, Jonathan. Generating English discourse from
semantic networks. Technical Report No. NL-3, National ScienceFoundation, Grant GJ 509 X, Computer Sciences Department andComputer-ssisted Instruction Laboratory, The University ofTexas at Austin, November, 1970b.
Simmons, R. L. Compiling semantic networks from English sentences.Unpublished manuscript, Computer Sciences Department, TheUniversity of Texas at Austin, 1971.
Zinn, Karl L. Instructional use of interactive computer systems.Datanation, September, 1968.
59